Carnegie Hall Data Lab
Introducing Stardog Voicebox
Carnegie Hall is embarking on a unique partnership with a company called Stardog, to help develop and test an artificial intelligence (AI) tool called Voicebox to support online visitors with searching and exploring Carnegie Hall’s performance history of nearly 63,000 events. We’re inviting visitors to the Carnegie Hall Data Lab site to help us with this exciting opportunity.
What are we asking you to do? We’d like you to spend a few minutes asking some fun questions about Carnegie Hall’s performance history and sharing some feedback on your experience and the quality of the answers you receive.
What is Stardog? Stardog is an enterprise knowledge graph platform that we use to publish Carnegie Hall’s performance history data in an open semantic data format.
What is Voicebox? Voicebox is a conversational AI chat interface for our performance history data.
Voicebox is a generative AI solution that answers your questions using Carnegie Hall’s connected performance history data in the Stardog knowledge graph. Unlike other commercial generative AI chatbots like ChatGPT, Voicebox does not use a large language model (LLM) to directly answer your questions. Instead Voicebox uses the LLMs to turn your questions into a structured SPARQL query over the knowledge graph. Voicebox learns how Carnegie Hall’s data is put together and interacts with the data set so that you don’t have to learn the SPARQL query language. This removes any opportunity for LLM hallucinations because the answers directly come from the stored data. Any hallucinations that might occur during query generation are caught and rectified because the queries are validated against Carnegie Hall’s data model.
Ask About Our History with Stardog Voicebox
You can find the Carnegie Hall Data Lab instance of Voicebox by following this link.
Tips for talking to Voicebox:
- Be as specific as possible
- For instance, “List the top 10 most performed works” is better than “Top 10 works”
- Use structured, grammatically correct requests
- For instance, “How many world premieres are given each year at Carnegie Hall” is better than “World premieres by year”
- This is a closed network of information meaning that the AI can only answer questions based on Carnegie Hall's performance history data and cannot interact with the open internet in any way to get more information.
Things Voicebox should be able to answer:
- Anything you would find on Carnegie Hall's Performance History Search
- Questions about composer and performer birthdays and birthplaces
- “How many”-type questions
- How many times has [x] performed at Carnegie Hall?
- How many rock concerts have there been at Carnegie Hall?
Things Voicebox cannot answer:
- Personal information about composers and performers besides birthdates and birthplaces
- Performance duration/run time
- Questions about gender. This is currently out of scope for this project.
How to provide feedback:
Email us: archives@carnegiehall.org
Please be kind! We’re continuing to make improvements to the model, and you might experience instances where Voicebox says it cannot answer your question, and you may receive incorrect responses. Voicebox is learning how to take your natural language questions and turn them into structured database queries. This feedback is good for us and part of the information we need to improve this new way to interact with our performance history data.